static int psmx_getinfo(uint32_t version, const char *node, const char *service,
uint64_t flags, struct fi_info *hints, struct fi_info **info)
{
struct fi_info *psmx_info;
uint32_t cnt = 0;
void *dest_addr = NULL;
int ep_type = FI_EP_RDM;
int av_type = FI_AV_UNSPEC;
enum fi_mr_mode mr_mode = FI_MR_SCALABLE;
int caps = 0;
uint64_t max_tag_value = 0;
int err = -FI_ENODATA;
FI_INFO(&psmx_prov, FI_LOG_CORE,"\n");
*info = NULL;
if (psm_ep_num_devunits(&cnt) || !cnt) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"no PSM device is found.\n");
return -FI_ENODATA;
}
if (node && !(flags & FI_SOURCE))
dest_addr = psmx_resolve_name(node, 0);
if (hints) {
switch (hints->addr_format) {
case FI_FORMAT_UNSPEC:
case FI_ADDR_PSMX:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->addr_format=%d, supported=%d,%d.\n",
hints->addr_format, FI_FORMAT_UNSPEC, FI_ADDR_PSMX);
goto err_out;
}
if (hints->ep_attr) {
switch (hints->ep_attr->type) {
case FI_EP_UNSPEC:
case FI_EP_DGRAM:
case FI_EP_RDM:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->type=%d, supported=%d,%d,%d.\n",
hints->ep_attr->type, FI_EP_UNSPEC,
FI_EP_DGRAM, FI_EP_RDM);
goto err_out;
}
switch (hints->ep_attr->protocol) {
case FI_PROTO_UNSPEC:
case FI_PROTO_PSMX:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->protocol=%d, supported=%d %d\n",
hints->ep_attr->protocol,
FI_PROTO_UNSPEC, FI_PROTO_PSMX);
goto err_out;
}
if (hints->ep_attr->tx_ctx_cnt > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->tx_ctx_cnt=%d, supported=0,1\n",
hints->ep_attr->tx_ctx_cnt);
goto err_out;
}
if (hints->ep_attr->rx_ctx_cnt > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->rx_ctx_cnt=%d, supported=0,1\n",
hints->ep_attr->rx_ctx_cnt);
goto err_out;
}
}
if ((hints->caps & PSMX_CAPS) != hints->caps &&
(hints->caps & PSMX_CAPS2) != hints->caps) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->caps=0x%llx, supported=0x%llx,0x%llx\n",
hints->caps, PSMX_CAPS, PSMX_CAPS2);
goto err_out;
}
if (hints->tx_attr) {
if ((hints->tx_attr->op_flags & PSMX_OP_FLAGS) !=
hints->tx_attr->op_flags) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx->flags=0x%llx, "
"supported=0x%llx\n",
hints->tx_attr->op_flags,
PSMX_OP_FLAGS);
goto err_out;
}
if (hints->tx_attr->inject_size > PSMX_INJECT_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->inject_size=%ld,"
"supported=%ld.\n",
hints->tx_attr->inject_size,
PSMX_INJECT_SIZE);
goto err_out;
}
}
if (hints->rx_attr &&
(hints->rx_attr->op_flags & PSMX_OP_FLAGS) !=
hints->rx_attr->op_flags) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->rx->flags=0x%llx, supported=0x%llx\n",
hints->rx_attr->op_flags, PSMX_OP_FLAGS);
goto err_out;
}
if ((hints->mode & PSMX_MODE) != PSMX_MODE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->mode=0x%llx, required=0x%llx\n",
hints->mode, PSMX_MODE);
goto err_out;
}
if (hints->fabric_attr && hints->fabric_attr->name &&
strncmp(hints->fabric_attr->name, PSMX_FABRIC_NAME, PSMX_FABRIC_NAME_LEN)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->fabric_name=%s, supported=psm\n",
hints->fabric_attr->name);
goto err_out;
}
if (hints->domain_attr) {
if (hints->domain_attr->name &&
strncmp(hints->domain_attr->name, PSMX_DOMAIN_NAME,
PSMX_DOMAIN_NAME_LEN)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_name=%s, supported=psm\n",
hints->domain_attr->name);
goto err_out;
}
switch (hints->domain_attr->av_type) {
case FI_AV_UNSPEC:
case FI_AV_MAP:
case FI_AV_TABLE:
av_type = hints->domain_attr->av_type;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->av_type=%d, supported=%d %d %d\n",
hints->domain_attr->av_type, FI_AV_UNSPEC, FI_AV_MAP,
FI_AV_TABLE);
goto err_out;
}
switch (hints->domain_attr->mr_mode) {
case FI_MR_UNSPEC:
break;
case FI_MR_BASIC:
case FI_MR_SCALABLE:
mr_mode = hints->domain_attr->mr_mode;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->mr_mode=%d, supported=%d %d %d\n",
hints->domain_attr->mr_mode, FI_MR_UNSPEC, FI_MR_BASIC,
FI_MR_SCALABLE);
goto err_out;
}
}
if (hints->ep_attr) {
if (hints->ep_attr->max_msg_size > PSMX_MAX_MSG_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->max_msg_size=%ld,"
"supported=%ld.\n",
hints->ep_attr->max_msg_size,
PSMX_MAX_MSG_SIZE);
goto err_out;
}
max_tag_value = fi_tag_bits(hints->ep_attr->mem_tag_format);
}
caps = hints->caps;
/* TODO: check other fields of hints */
}
if (psmx_reserve_tag_bits(&caps, &max_tag_value) < 0)
goto err_out;
psmx_info = fi_allocinfo();
if (!psmx_info) {
err = -FI_ENOMEM;
goto err_out;
}
psmx_info->ep_attr->type = ep_type;
psmx_info->ep_attr->protocol = FI_PROTO_PSMX;
psmx_info->ep_attr->protocol_version = PSM_VERNO;
psmx_info->ep_attr->max_msg_size = PSMX_MAX_MSG_SIZE;
psmx_info->ep_attr->mem_tag_format = fi_tag_format(max_tag_value);
psmx_info->ep_attr->tx_ctx_cnt = 1;
psmx_info->ep_attr->rx_ctx_cnt = 1;
psmx_info->domain_attr->threading = FI_THREAD_COMPLETION;
psmx_info->domain_attr->control_progress = FI_PROGRESS_MANUAL;
psmx_info->domain_attr->data_progress = FI_PROGRESS_MANUAL;
psmx_info->domain_attr->name = strdup(PSMX_DOMAIN_NAME);
psmx_info->domain_attr->resource_mgmt = FI_RM_ENABLED;
psmx_info->domain_attr->av_type = av_type;
psmx_info->domain_attr->mr_mode = mr_mode;
psmx_info->domain_attr->mr_key_size = sizeof(uint64_t);
psmx_info->domain_attr->cq_data_size = 4;
psmx_info->domain_attr->cq_cnt = 65535;
psmx_info->domain_attr->ep_cnt = 65535;
psmx_info->domain_attr->tx_ctx_cnt = 1;
psmx_info->domain_attr->rx_ctx_cnt = 1;
psmx_info->domain_attr->max_ep_tx_ctx = 65535;
psmx_info->domain_attr->max_ep_rx_ctx = 1;
psmx_info->domain_attr->max_ep_stx_ctx = 65535;
psmx_info->domain_attr->max_ep_srx_ctx = 0;
psmx_info->next = NULL;
psmx_info->caps = (hints && hints->caps) ? hints->caps : caps;
psmx_info->mode = PSMX_MODE;
psmx_info->addr_format = FI_ADDR_PSMX;
psmx_info->src_addrlen = 0;
psmx_info->dest_addrlen = sizeof(psm_epid_t);
psmx_info->src_addr = NULL;
psmx_info->dest_addr = dest_addr;
psmx_info->fabric_attr->name = strdup(PSMX_FABRIC_NAME);
psmx_info->fabric_attr->prov_name = strdup(PSMX_PROV_NAME);
psmx_info->tx_attr->caps = psmx_info->caps;
psmx_info->tx_attr->mode = psmx_info->mode;
psmx_info->tx_attr->op_flags = (hints && hints->tx_attr && hints->tx_attr->op_flags)
? hints->tx_attr->op_flags : 0;
psmx_info->tx_attr->msg_order = FI_ORDER_SAS;
psmx_info->tx_attr->comp_order = FI_ORDER_NONE;
psmx_info->tx_attr->inject_size = PSMX_INJECT_SIZE;
psmx_info->tx_attr->size = UINT64_MAX;
psmx_info->tx_attr->iov_limit = 1;
psmx_info->rx_attr->caps = psmx_info->caps;
psmx_info->rx_attr->mode = psmx_info->mode;
psmx_info->rx_attr->op_flags = (hints && hints->rx_attr && hints->tx_attr->op_flags)
? hints->tx_attr->op_flags : 0;
psmx_info->rx_attr->msg_order = FI_ORDER_SAS;
psmx_info->rx_attr->comp_order = FI_ORDER_NONE;
psmx_info->rx_attr->total_buffered_recv = ~(0ULL); /* that's how PSM handles it internally! */
psmx_info->rx_attr->size = UINT64_MAX;
psmx_info->rx_attr->iov_limit = 1;
*info = psmx_info;
return 0;
err_out:
return err;
}
psm_error_t
__psm_ep_open_internal(psm_uuid_t const unique_job_key, int *devid_enabled,
struct psm_ep_open_opts const *opts_i, psm_mq_t mq,
psm_ep_t *epo, psm_epid_t *epido)
{
psm_ep_t ep = NULL;
uint32_t num_units;
size_t len;
psm_error_t err;
psm_epaddr_t epaddr = NULL;
char buf[128], *p, *e;
union psmi_envvar_val envvar_val;
size_t ptl_sizes;
struct psm_ep_open_opts opts;
ptl_t *amsh_ptl, *ips_ptl, *self_ptl;
int i;
/* First get the set of default options, we overwrite with the user's
* desired values afterwards */
if ((err = psm_ep_open_opts_get_defaults(&opts)))
goto fail;
if (opts_i != NULL) {
if (opts_i->timeout != -1)
opts.timeout = opts_i->timeout;
if (opts_i->unit != -1)
opts.unit = opts_i->unit;
if (opts_i->affinity != -1)
opts.affinity = opts_i->affinity;
if (opts_i->sendbufs_num != -1)
opts.sendbufs_num = opts_i->sendbufs_num;
if (opts_i->network_pkey != HFI_DEFAULT_P_KEY)
opts.network_pkey = opts_i->network_pkey;
if (opts_i->port != 0)
opts.port = opts_i->port;
if (opts_i->outsl != -1)
opts.outsl = opts_i->outsl;
if (opts_i->service_id)
opts.service_id = (uint64_t) opts_i->service_id;
if (opts_i->path_res_type != PSM_PATH_RES_NONE)
opts.path_res_type = opts_i->path_res_type;
if (opts_i->senddesc_num)
opts.senddesc_num = opts_i->senddesc_num;
if (opts_i->imm_size)
opts.imm_size = opts_i->imm_size;
}
/* Get Service ID from environment */
if (!psmi_getenv("PSM_IB_SERVICE_ID",
"HFI Service ID for path resolution",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_ULONG_ULONG,
(union psmi_envvar_val)HFI_DEFAULT_SERVICE_ID,
&envvar_val)) {
opts.service_id = (uint64_t) envvar_val.e_ulonglong;
}
/* Get Path resolution type from environment Possible choices are:
*
* NONE : Default same as previous instances. Utilizes static data.
* OPP : Use OFED Plus Plus library to do path record queries.
* UMAD : Use raw libibumad interface to form and process path records.
*/
if (!psmi_getenv("PSM_PATH_REC",
"Mechanism to query HFI path record (default is no path query)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_STR,
(union psmi_envvar_val)"none", &envvar_val)) {
if (!strcasecmp(envvar_val.e_str, "none"))
opts.path_res_type = PSM_PATH_RES_NONE;
else if (!strcasecmp(envvar_val.e_str, "opp"))
opts.path_res_type = PSM_PATH_RES_OPP;
else if (!strcasecmp(envvar_val.e_str, "umad"))
opts.path_res_type = PSM_PATH_RES_UMAD;
else {
_HFI_ERROR("Unknown path resolution type %s. "
"Disabling use of path record query.\n",
envvar_val.e_str);
opts.path_res_type = PSM_PATH_RES_NONE;
}
}
/* If a specific unit is set in the environment, use that one. */
if (!psmi_getenv("HFI_UNIT", "Device Unit number (-1 autodetects)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_LONG,
(union psmi_envvar_val)HFI_UNIT_ID_ANY, &envvar_val)) {
opts.unit = envvar_val.e_long;
}
/* Get user specified port number to use. */
if (!psmi_getenv("HFI_PORT", "IB Port number (0 autodetects)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_LONG,
(union psmi_envvar_val)HFI_PORT_NUM_ANY,
&envvar_val)) {
opts.port = envvar_val.e_long;
}
/* Get service level from environment, path-query overrides it */
if (!psmi_getenv
("HFI_SL", "HFI outging ServiceLevel number (default 0)",
PSMI_ENVVAR_LEVEL_USER, PSMI_ENVVAR_TYPE_LONG,
(union psmi_envvar_val)PSMI_SL_DEFAULT, &envvar_val)) {
opts.outsl = envvar_val.e_long;
}
/* Get network key from environment. MVAPICH and other vendor MPIs do not
* specify it on ep open and we may require it for vFabrics.
* path-query will override it.
*/
if (!psmi_getenv("PSM_PKEY",
"HFI PKey to use for endpoint",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_ULONG,
(union psmi_envvar_val)HFI_DEFAULT_P_KEY,
&envvar_val)) {
opts.network_pkey = (uint64_t) envvar_val.e_ulong;
}
/* BACKWARDS COMPATIBILITY: Open MPI likes to choose its own PKEY of
0x7FFF. That's no longer a valid default, so override it if the
client was compiled against PSM v1 */
if (PSMI_VERNO_GET_MAJOR(psmi_verno_client()) < 2 &&
opts.network_pkey == 0x7FFF) {
opts.network_pkey = HFI_DEFAULT_P_KEY;
}
/* Get number of default send buffers from environment */
if (!psmi_getenv("PSM_NUM_SEND_BUFFERS",
"Number of send buffers to allocate [1024]",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)1024, &envvar_val)) {
opts.sendbufs_num = envvar_val.e_uint;
}
/* Get immediate data size - transfers less than immediate data size do
* not consume a send buffer and require just a send descriptor.
*/
if (!psmi_getenv("PSM_SEND_IMMEDIATE_SIZE",
"Immediate data send size not requiring a buffer [128]",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)128, &envvar_val)) {
opts.imm_size = envvar_val.e_uint;
}
/* Get numner of send descriptors - by default this is 4 times the number
* of send buffers - mainly used for short/inlined messages.
*/
if (!psmi_getenv("PSM_NUM_SEND_DESCRIPTORS",
"Number of send descriptors to allocate [4096]",
PSMI_ENVVAR_LEVEL_USER,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)4096, &envvar_val)) {
opts.senddesc_num = envvar_val.e_uint;
}
if (psmi_device_is_enabled(devid_enabled, PTL_DEVID_IPS)) {
if ((err = psm_ep_num_devunits(&num_units)) != PSM_OK)
goto fail;
} else
num_units = 0;
/* do some error checking */
if (opts.timeout < -1) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid timeout value %lld",
(long long)opts.timeout);
goto fail;
} else if (num_units && (opts.unit < -1 || opts.unit >= (int)num_units)) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid Device Unit ID %d (%d units found)",
opts.unit, num_units);
goto fail;
} else if (opts.port < 0 || opts.port > HFI_MAX_PORT) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid Device port number %d",
opts.port);
goto fail;
} else if (opts.affinity < 0
|| opts.affinity > PSM_EP_OPEN_AFFINITY_FORCE) {
err =
psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid Affinity option: %d",
opts.affinity);
goto fail;
} else if (opts.outsl < PSMI_SL_MIN || opts.outsl > PSMI_SL_MAX) {
err = psmi_handle_error(NULL, PSM_PARAM_ERR,
"Invalid SL number: %lld",
(unsigned long long)opts.outsl);
goto fail;
}
/* Set environment variable if PSM is not allowed to set affinity */
if (opts.affinity == PSM_EP_OPEN_AFFINITY_SKIP)
setenv("HFI_NO_CPUAFFINITY", "1", 1);
/* Allocate end point structure storage */
ptl_sizes =
(psmi_device_is_enabled(devid_enabled, PTL_DEVID_SELF) ?
psmi_ptl_self.sizeof_ptl() : 0) +
(psmi_device_is_enabled(devid_enabled, PTL_DEVID_IPS) ?
psmi_ptl_ips.sizeof_ptl() : 0) +
(psmi_device_is_enabled(devid_enabled, PTL_DEVID_AMSH) ?
psmi_ptl_amsh.sizeof_ptl() : 0);
if (ptl_sizes == 0)
return PSM_EP_NO_DEVICE;
ep = (psm_ep_t) psmi_memalign(PSMI_EP_NONE, UNDEFINED, 64,
sizeof(struct psm_ep) + ptl_sizes);
epaddr = (psm_epaddr_t) psmi_calloc(PSMI_EP_NONE, PER_PEER_ENDPOINT,
1, sizeof(struct psm_epaddr));
if (ep == NULL || epaddr == NULL) {
err = psmi_handle_error(NULL, PSM_NO_MEMORY,
"Couldn't allocate memory for %s structure",
ep == NULL ? "psm_ep" : "psm_epaddr");
goto fail;
}
/* Copy PTL enabled status */
for (i = 0; i < PTL_MAX_INIT; i++)
ep->devid_enabled[i] = devid_enabled[i];
/* Matched Queue initialization. We do this early because we have to
* make sure ep->mq exists and is valid before calling ips_do_work.
*/
ep->mq = mq;
/* Get ready for PTL initialization */
memcpy(&ep->uuid, (void *)unique_job_key, sizeof(psm_uuid_t));
ep->epaddr = epaddr;
ep->memmode = mq->memmode;
ep->hfi_num_sendbufs = opts.sendbufs_num;
ep->service_id = opts.service_id;
ep->path_res_type = opts.path_res_type;
ep->hfi_num_descriptors = opts.senddesc_num;
ep->hfi_imm_size = opts.imm_size;
ep->errh = psmi_errhandler_global; /* by default use the global one */
ep->ptl_amsh.ep_poll = psmi_poll_noop;
ep->ptl_ips.ep_poll = psmi_poll_noop;
ep->connections = 0;
/* See how many iterations we want to spin before yielding */
psmi_getenv("PSM_YIELD_SPIN_COUNT",
"Spin poll iterations before yield",
PSMI_ENVVAR_LEVEL_HIDDEN,
PSMI_ENVVAR_TYPE_UINT,
(union psmi_envvar_val)PSMI_BLOCKUNTIL_POLLS_BEFORE_YIELD,
&envvar_val);
ep->yield_spin_cnt = envvar_val.e_uint;
ptl_sizes = 0;
amsh_ptl = ips_ptl = self_ptl = NULL;
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH)) {
amsh_ptl = (ptl_t *) (ep->ptl_base_data + ptl_sizes);
ptl_sizes += psmi_ptl_amsh.sizeof_ptl();
}
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS)) {
ips_ptl = (ptl_t *) (ep->ptl_base_data + ptl_sizes);
ptl_sizes += psmi_ptl_ips.sizeof_ptl();
}
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_SELF)) {
self_ptl = (ptl_t *) (ep->ptl_base_data + ptl_sizes);
ptl_sizes += psmi_ptl_self.sizeof_ptl();
}
if ((err = psmi_ep_open_device(ep, &opts, unique_job_key,
&(ep->context), &ep->epid)))
goto fail;
psmi_assert_always(ep->epid != 0);
ep->epaddr->epid = ep->epid;
_HFI_VDBG("psmi_ep_open_device() passed\n");
/* Set our new label as soon as we know what it is */
strncpy(buf, psmi_gethostname(), sizeof(buf) - 1);
buf[sizeof(buf) - 1] = '\0';
p = buf + strlen(buf);
/* If our rank is set, use it. If not, use context.subcontext notation */
if (((e = getenv("MPI_RANKID")) != NULL && *e) ||
((e = getenv("PSC_MPI_RANK")) != NULL && *e))
len = snprintf(p, sizeof(buf) - strlen(buf), ":%d.", atoi(e));
else
len = snprintf(p, sizeof(buf) - strlen(buf), ":%d.%d.",
(uint32_t) psm_epid_context(ep->epid),
(uint32_t) psmi_epid_subcontext(ep->epid));
*(p + len) = '\0';
ep->context_mylabel = psmi_strdup(ep, buf);
if (ep->context_mylabel == NULL) {
err = PSM_NO_MEMORY;
goto fail;
}
/* hfi_set_mylabel(ep->context_mylabel); */
if ((err = psmi_epid_set_hostname(psm_epid_nid(ep->epid), buf, 0)))
goto fail;
_HFI_VDBG("start ptl device init...\n");
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_SELF)) {
if ((err = psmi_ptl_self.init(ep, self_ptl, &ep->ptl_self)))
goto fail;
}
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_IPS)) {
if ((err = psmi_ptl_ips.init(ep, ips_ptl, &ep->ptl_ips)))
goto fail;
}
/* If we're shm-only, this device is enabled above */
if (psmi_ep_device_is_enabled(ep, PTL_DEVID_AMSH)) {
if ((err = psmi_ptl_amsh.init(ep, amsh_ptl, &ep->ptl_amsh)))
goto fail;
} else {
/* We may have pre-attached as part of getting our rank for enabling
* shared contexts. */
}
_HFI_VDBG("finish ptl device init...\n");
/*
* Keep only IPS since only IPS support multi-rail, other devices
* are only setup once. IPS device can come to this function again.
*/
for (i = 0; i < PTL_MAX_INIT; i++) {
if (devid_enabled[i] != PTL_DEVID_IPS) {
devid_enabled[i] = -1;
}
}
*epido = ep->epid;
*epo = ep;
return PSM_OK;
fail:
if (ep != NULL) {
if (ep->context.fd != -1)
close(ep->context.fd);
psmi_free(ep);
}
if (epaddr != NULL)
psmi_free(epaddr);
return err;
}
static psm_error_t
psmi_ep_devlids(uint16_t **lids, uint32_t *num_lids_o,
uint64_t my_gid_hi, uint64_t my_gid_lo)
{
static uint16_t *hfi_lids;
static uint32_t nlids;
uint32_t num_units;
int i;
psm_error_t err = PSM_OK;
PSMI_ERR_UNLESS_INITIALIZED(NULL);
if (hfi_lids == NULL) {
if ((err = psm_ep_num_devunits(&num_units)))
goto fail;
hfi_lids = (uint16_t *)
psmi_calloc(PSMI_EP_NONE, UNDEFINED,
num_units * HFI_MAX_PORT, sizeof(uint16_t));
if (hfi_lids == NULL) {
err = psmi_handle_error(NULL, PSM_NO_MEMORY,
"Couldn't allocate memory for dev_lids structure");
goto fail;
}
for (i = 0; i < num_units; i++) {
int j;
for (j = 1; j <= HFI_MAX_PORT; j++) {
int lid = hfi_get_port_lid(i, j);
int ret;
uint64_t gid_hi = 0, gid_lo = 0;
if (lid == -1)
continue;
ret = hfi_get_port_gid(i, j, &gid_hi, &gid_lo);
if (ret == -1)
continue;
else if (my_gid_hi != gid_hi) {
_HFI_VDBG("LID %d, unit %d, port %d, "
"mismatched GID %llx:%llx and "
"%llx:%llx\n",
lid, i, j,
(unsigned long long)gid_hi,
(unsigned long long)gid_lo,
(unsigned long long)my_gid_hi,
(unsigned long long)
my_gid_lo);
continue;
}
_HFI_VDBG("LID %d, unit %d, port %d, "
"matching GID %llx:%llx and "
"%llx:%llx\n", lid, i, j,
(unsigned long long)gid_hi,
(unsigned long long)gid_lo,
(unsigned long long)my_gid_hi,
(unsigned long long)my_gid_lo);
hfi_lids[nlids++] = (uint16_t) lid;
}
}
if (nlids == 0) {
err = psmi_handle_error(NULL, PSM_EP_DEVICE_FAILURE,
"Couldn't get lid&gid from any unit/port");
goto fail;
}
}
*lids = hfi_lids;
*num_lids_o = nlids;
fail:
return err;
}
static psm_error_t
psmi_ep_multirail(int *num_rails, uint32_t *unit, uint16_t *port)
{
uint32_t num_units;
uint64_t gid_hi, gid_lo;
int i, j, ret, count = 0;
char *env;
psm_error_t err = PSM_OK;
uint64_t gidh[HFI_MAX_RAILS][3];
env = getenv("PSM_MULTIRAIL");
if (!env || atoi(env) == 0) {
*num_rails = 0;
return err;
}
/*
* map is in format: unit:port,unit:port,...
*/
if ((env = getenv("PSM_MULTIRAIL_MAP"))) {
if (sscanf(env, "%d:%d", &i, &j) == 2) {
char *comma = strchr(env, ',');
unit[count] = i;
port[count] = j;
count++;
while (comma) {
if (sscanf(comma, ",%d:%d", &i, &j) != 2) {
break;
}
unit[count] = i;
port[count] = j;
count++;
if (count == HFI_MAX_RAILS)
break;
comma = strchr(comma + 1, ',');
}
}
*num_rails = count;
/*
* Check if any of the port is not usable.
*/
for (i = 0; i < count; i++) {
ret = hfi_get_port_lid(unit[i], port[i]);
if (ret == -1) {
err =
psmi_handle_error(NULL,
PSM_EP_DEVICE_FAILURE,
"Couldn't get lid for unit %d:%d",
unit[i], port[i]);
return err;
}
ret =
hfi_get_port_gid(unit[i], port[i], &gid_hi,
&gid_lo);
if (ret == -1) {
err =
psmi_handle_error(NULL,
PSM_EP_DEVICE_FAILURE,
"Couldn't get gid for unit %d:%d",
unit[i], port[i]);
return err;
}
}
return err;
}
if ((err = psm_ep_num_devunits(&num_units))) {
return err;
}
if (num_units > HFI_MAX_RAILS) {
_HFI_INFO
("Found %d units, max %d units are supported, use %d\n",
num_units, HFI_MAX_RAILS, HFI_MAX_RAILS);
num_units = HFI_MAX_RAILS;
}
/*
* Get all the ports with a valid lid and gid, one per unit.
*/
for (i = 0; i < num_units; i++) {
for (j = 1; j <= HFI_MAX_PORT; j++) {
ret = hfi_get_port_lid(i, j);
if (ret == -1)
continue;
ret = hfi_get_port_gid(i, j, &gid_hi, &gid_lo);
if (ret == -1)
continue;
gidh[count][0] = gid_hi;
gidh[count][1] = i;
gidh[count][2] = j;
count++;
break;
}
}
/*
* Sort all the ports with gidh from small to big.
* This is for multiple fabrics, and we use fabric with the
* smallest gid to make the master connection.
*/
qsort(gidh, count, sizeof(uint64_t) * 3, cmpfunc);
for (i = 0; i < count; i++) {
unit[i] = (uint32_t) gidh[i][1];
port[i] = (uint16_t) (uint32_t) gidh[i][2];
}
*num_rails = count;
return err;
}
static int psmx_getinfo(uint32_t version, const char *node, const char *service,
uint64_t flags, const struct fi_info *hints,
struct fi_info **info)
{
struct fi_info *psmx_info;
uint32_t cnt = 0;
psm_epid_t *dest_addr = NULL;
struct psmx_src_name *src_addr = NULL;
int ep_type = FI_EP_RDM;
int av_type = FI_AV_UNSPEC;
uint64_t mode = FI_CONTEXT;
enum fi_mr_mode mr_mode = FI_MR_SCALABLE;
enum fi_threading threading = FI_THREAD_COMPLETION;
enum fi_progress control_progress = FI_PROGRESS_MANUAL;
enum fi_progress data_progress = FI_PROGRESS_MANUAL;
int caps = 0;
uint64_t max_tag_value = 0;
int err = -FI_ENODATA;
int svc0, svc = PSMX_ANY_SERVICE;
FI_INFO(&psmx_prov, FI_LOG_CORE,"\n");
*info = NULL;
/* Perform some quick check first to avoid unnecessary operations */
if (hints) {
if (hints->fabric_attr && hints->fabric_attr->name &&
strcasecmp(hints->fabric_attr->name, PSMX_FABRIC_NAME)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->fabric_name=%s, supported=%s\n",
hints->fabric_attr->name, PSMX_FABRIC_NAME);
goto err_out;
}
if (hints->domain_attr && hints->domain_attr->name &&
strcasecmp(hints->domain_attr->name, PSMX_DOMAIN_NAME)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_name=%s, supported=%s\n",
hints->domain_attr->name, PSMX_DOMAIN_NAME);
goto err_out;
}
if (hints->ep_attr) {
switch (hints->ep_attr->type) {
case FI_EP_UNSPEC:
case FI_EP_DGRAM:
case FI_EP_RDM:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->type=%d, supported=%d,%d,%d.\n",
hints->ep_attr->type, FI_EP_UNSPEC,
FI_EP_DGRAM, FI_EP_RDM);
goto err_out;
}
}
if ((hints->caps & PSMX_CAPS) != hints->caps &&
(hints->caps & PSMX_CAPS2) != hints->caps) {
uint64_t psmx_caps = PSMX_CAPS;
uint64_t psmx_caps2 = PSMX_CAPS2;
PSMX_INFO_DIFF("hints->caps", hints->caps, psmx_caps,
FI_TYPE_CAPS);
PSMX_INFO_DIFF("alternatively, hints->caps",
hints->caps, psmx_caps2, FI_TYPE_CAPS);
goto err_out;
}
}
if (FI_VERSION_GE(version, FI_VERSION(1,5)))
mr_mode = 0;
if (psmx_init_lib())
return -FI_ENODATA;
if (psmx_compat_lib) {
/*
* native PSM running over TrueScale doesn't have the issue handled
* here. it's only present when PSM is supported via the psm2-compat
* library, where the PSM functions are just wrappers around the PSM2
* counterparts.
*
* psm2_ep_num_devunits() may wait for 15 seconds before return
* when /dev/hfi1_0 is not present. Check the existence of any hfi1
* device interface first to avoid this delay. Note that the devices
* don't necessarily appear consecutively so we need to check all
* possible device names before returning "no device found" error.
* This also means if "/dev/hfi1_0" doesn't exist but other devices
* exist, we are still going to see the delay; but that's a rare case.
*/
glob_t glob_buf;
if ((glob("/dev/hfi1_[0-9]", 0, NULL, &glob_buf) != 0) &&
(glob("/dev/hfi1_[0-9][0-9]", GLOB_APPEND, NULL, &glob_buf) != 0)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"no hfi1 device is found.\n");
return -FI_ENODATA;
}
globfree(&glob_buf);
}
if (psm_ep_num_devunits(&cnt) || !cnt) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"no PSM device is found.\n");
return -FI_ENODATA;
}
src_addr = calloc(1, sizeof(*src_addr));
if (!src_addr) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"failed to allocate src addr.\n");
return -FI_ENODATA;
}
src_addr->signature = 0xFFFF;
src_addr->unit = PSMX_DEFAULT_UNIT;
src_addr->port = PSMX_DEFAULT_PORT;
src_addr->service = PSMX_ANY_SERVICE;
if (flags & FI_SOURCE) {
if (node)
sscanf(node, "%*[^:]:%" SCNi8 ":%" SCNu8, &src_addr->unit, &src_addr->port);
if (service)
sscanf(service, "%" SCNu32, &src_addr->service);
FI_INFO(&psmx_prov, FI_LOG_CORE,
"node '%s' service '%s' converted to <unit=%d, port=%d, service=%d>\n",
node, service, src_addr->unit, src_addr->port, src_addr->service);
} else if (node) {
psm_uuid_t uuid;
psmx_get_uuid(uuid);
struct util_ns ns = {
.port = psmx_uuid_to_port(uuid),
.name_len = sizeof(*dest_addr),
.service_len = sizeof(svc),
.service_cmp = psmx_ns_service_cmp,
.is_service_wildcard = psmx_ns_is_service_wildcard,
};
ofi_ns_init(&ns);
if (service)
svc = atoi(service);
svc0 = svc;
dest_addr = (psm_epid_t *)ofi_ns_resolve_name(&ns, node, &svc);
if (dest_addr) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"'%s:%u' resolved to <epid=%"PRIu64">:%u\n",
node, svc0, *dest_addr, svc);
} else {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"failed to resolve '%s:%u'.\n", node, svc);
err = -FI_ENODATA;
goto err_out;
}
}
if (hints) {
switch (hints->addr_format) {
case FI_FORMAT_UNSPEC:
case FI_ADDR_PSMX:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->addr_format=%d, supported=%d,%d.\n",
hints->addr_format, FI_FORMAT_UNSPEC, FI_ADDR_PSMX);
goto err_out;
}
if (hints->ep_attr) {
switch (hints->ep_attr->protocol) {
case FI_PROTO_UNSPEC:
case FI_PROTO_PSMX:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->protocol=%d, supported=%d %d\n",
hints->ep_attr->protocol,
FI_PROTO_UNSPEC, FI_PROTO_PSMX);
goto err_out;
}
if (hints->ep_attr->tx_ctx_cnt > 1 &&
hints->ep_attr->tx_ctx_cnt != FI_SHARED_CONTEXT) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->tx_ctx_cnt=%"PRIu64", supported=0,1\n",
hints->ep_attr->tx_ctx_cnt);
goto err_out;
}
if (hints->ep_attr->rx_ctx_cnt > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->rx_ctx_cnt=%"PRIu64", supported=0,1\n",
hints->ep_attr->rx_ctx_cnt);
goto err_out;
}
}
if (hints->tx_attr) {
if ((hints->tx_attr->op_flags & PSMX_OP_FLAGS) !=
hints->tx_attr->op_flags) {
uint64_t psmx_op_flags = PSMX_OP_FLAGS;
PSMX_INFO_DIFF("hints->tx_attr->of_flags",
hints->tx_attr->op_flags,
psmx_op_flags, FI_TYPE_OP_FLAGS);
goto err_out;
}
if (hints->tx_attr->inject_size > PSMX_INJECT_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->inject_size=%"PRIu64","
"supported=%d.\n",
hints->tx_attr->inject_size,
PSMX_INJECT_SIZE);
goto err_out;
}
}
if (hints->rx_attr &&
(hints->rx_attr->op_flags & PSMX_OP_FLAGS) !=
hints->rx_attr->op_flags) {
uint64_t psmx_op_flags = PSMX_OP_FLAGS;
PSMX_INFO_DIFF("hints->rx_attr->of_flags",
hints->rx_attr->op_flags,
psmx_op_flags, FI_TYPE_OP_FLAGS);
goto err_out;
}
if ((hints->caps & FI_TAGGED) ||
((hints->caps & FI_MSG) && !psmx_env.am_msg)) {
if ((hints->mode & FI_CONTEXT) != FI_CONTEXT) {
uint64_t psmx_mode = FI_CONTEXT;
PSMX_INFO_DIFF("hints->mode", hints->mode,
psmx_mode, FI_TYPE_MODE);
goto err_out;
}
} else {
mode = 0;
}
if (hints->domain_attr) {
switch (hints->domain_attr->av_type) {
case FI_AV_UNSPEC:
case FI_AV_MAP:
case FI_AV_TABLE:
av_type = hints->domain_attr->av_type;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->av_type=%d, supported=%d %d %d\n",
hints->domain_attr->av_type, FI_AV_UNSPEC, FI_AV_MAP,
FI_AV_TABLE);
goto err_out;
}
if (hints->domain_attr->mr_mode == FI_MR_BASIC) {
mr_mode = FI_MR_BASIC;
} else if (hints->domain_attr->mr_mode == FI_MR_SCALABLE) {
mr_mode = FI_MR_SCALABLE;
} else if (hints->domain_attr->mr_mode & (FI_MR_BASIC | FI_MR_SCALABLE)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->mr_mode has FI_MR_BASIC or FI_MR_SCALABLE "
"combined with other bits\n");
goto err_out;
}
switch (hints->domain_attr->threading) {
case FI_THREAD_UNSPEC:
break;
case FI_THREAD_FID:
case FI_THREAD_ENDPOINT:
case FI_THREAD_COMPLETION:
case FI_THREAD_DOMAIN:
threading = hints->domain_attr->threading;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->threading=%d, supported=%d %d %d %d %d\n",
hints->domain_attr->threading, FI_THREAD_UNSPEC,
FI_THREAD_FID, FI_THREAD_ENDPOINT, FI_THREAD_COMPLETION,
FI_THREAD_DOMAIN);
goto err_out;
}
switch (hints->domain_attr->control_progress) {
case FI_PROGRESS_UNSPEC:
break;
case FI_PROGRESS_MANUAL:
case FI_PROGRESS_AUTO:
control_progress = hints->domain_attr->control_progress;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->control_progress=%d, supported=%d %d %d\n",
hints->domain_attr->control_progress, FI_PROGRESS_UNSPEC,
FI_PROGRESS_MANUAL, FI_PROGRESS_AUTO);
goto err_out;
}
switch (hints->domain_attr->data_progress) {
case FI_PROGRESS_UNSPEC:
break;
case FI_PROGRESS_MANUAL:
case FI_PROGRESS_AUTO:
data_progress = hints->domain_attr->data_progress;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->data_progress=%d, supported=%d %d %d\n",
hints->domain_attr->data_progress, FI_PROGRESS_UNSPEC,
FI_PROGRESS_MANUAL, FI_PROGRESS_AUTO);
goto err_out;
}
if (hints->domain_attr->caps & FI_SHARED_AV) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->caps=%lx, shared AV is unsupported\n",
hints->domain_attr->caps);
goto err_out;
}
}
if (hints->ep_attr) {
if (hints->ep_attr->max_msg_size > PSMX_MAX_MSG_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->max_msg_size=%"PRIu64","
"supported=%llu.\n",
hints->ep_attr->max_msg_size,
PSMX_MAX_MSG_SIZE);
goto err_out;
}
max_tag_value = ofi_max_tag(hints->ep_attr->mem_tag_format);
}
if (hints->tx_attr) {
if ((hints->tx_attr->msg_order & PSMX_MSG_ORDER) !=
hints->tx_attr->msg_order) {
uint64_t psmx_msg_order = PSMX_MSG_ORDER;
PSMX_INFO_DIFF("hints->tx_attr->msg_order",
hints->tx_attr->msg_order,
psmx_msg_order, FI_TYPE_MSG_ORDER);
goto err_out;
}
if ((hints->tx_attr->comp_order & PSMX_COMP_ORDER) !=
hints->tx_attr->comp_order) {
uint64_t psmx_comp_order = PSMX_COMP_ORDER;
PSMX_INFO_DIFF("hints->tx_attr->comp_order",
hints->tx_attr->comp_order,
psmx_comp_order, FI_TYPE_MSG_ORDER);
goto err_out;
}
if (hints->tx_attr->inject_size > PSMX_INJECT_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->inject_size=%ld,"
"supported=%d.\n",
hints->tx_attr->inject_size,
PSMX_INJECT_SIZE);
goto err_out;
}
if (hints->tx_attr->iov_limit > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->iov_limit=%ld,"
"supported=1.\n",
hints->tx_attr->iov_limit);
goto err_out;
}
if (hints->tx_attr->rma_iov_limit > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->rma_iov_limit=%ld,"
"supported=1.\n",
hints->tx_attr->rma_iov_limit);
goto err_out;
}
}
if (hints->rx_attr) {
if ((hints->rx_attr->msg_order & PSMX_MSG_ORDER) !=
hints->rx_attr->msg_order) {
uint64_t psmx_msg_order = PSMX_MSG_ORDER;
PSMX_INFO_DIFF("hints->rx_attr->msg_order",
hints->rx_attr->msg_order,
psmx_msg_order, FI_TYPE_MSG_ORDER);
goto err_out;
}
if ((hints->rx_attr->comp_order & PSMX_COMP_ORDER) !=
hints->rx_attr->comp_order) {
uint64_t psmx_comp_order = PSMX_COMP_ORDER;
PSMX_INFO_DIFF("hints->rx_attr->comp_order",
hints->rx_attr->comp_order,
psmx_comp_order, FI_TYPE_MSG_ORDER);
goto err_out;
}
if (hints->rx_attr->iov_limit > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->rx_attr->iov_limit=%ld,"
"supported=1.\n",
hints->rx_attr->iov_limit);
goto err_out;
}
}
caps = hints->caps;
/* TODO: check other fields of hints */
}
if (psmx_reserve_tag_bits(&caps, &max_tag_value) < 0)
goto err_out;
psmx_info = fi_allocinfo();
if (!psmx_info) {
err = -FI_ENOMEM;
goto err_out;
}
psmx_info->ep_attr->type = ep_type;
psmx_info->ep_attr->protocol = FI_PROTO_PSMX;
psmx_info->ep_attr->protocol_version = PSM_VERNO;
psmx_info->ep_attr->max_msg_size = PSMX_MAX_MSG_SIZE;
psmx_info->ep_attr->max_order_raw_size = PSMX_RMA_ORDER_SIZE;
psmx_info->ep_attr->max_order_war_size = PSMX_RMA_ORDER_SIZE;
psmx_info->ep_attr->max_order_waw_size = PSMX_RMA_ORDER_SIZE;
psmx_info->ep_attr->mem_tag_format = ofi_tag_format(max_tag_value);
psmx_info->ep_attr->tx_ctx_cnt = 1;
psmx_info->ep_attr->rx_ctx_cnt = 1;
psmx_info->domain_attr->threading = threading;
psmx_info->domain_attr->control_progress = control_progress;
psmx_info->domain_attr->data_progress = data_progress;
psmx_info->domain_attr->name = strdup(PSMX_DOMAIN_NAME);
psmx_info->domain_attr->resource_mgmt = FI_RM_ENABLED;
psmx_info->domain_attr->av_type = av_type;
psmx_info->domain_attr->mr_mode = mr_mode;
psmx_info->domain_attr->mr_key_size = sizeof(uint64_t);
psmx_info->domain_attr->cq_data_size = 4;
psmx_info->domain_attr->cq_cnt = 65535;
psmx_info->domain_attr->ep_cnt = 65535;
psmx_info->domain_attr->tx_ctx_cnt = 1;
psmx_info->domain_attr->rx_ctx_cnt = 1;
psmx_info->domain_attr->max_ep_tx_ctx = 1;
psmx_info->domain_attr->max_ep_rx_ctx = 1;
psmx_info->domain_attr->max_ep_stx_ctx = 65535;
psmx_info->domain_attr->max_ep_srx_ctx = 0;
psmx_info->domain_attr->cntr_cnt = 65535;
psmx_info->domain_attr->mr_iov_limit = 65535;
psmx_info->domain_attr->caps = PSMX_DOM_CAPS;
psmx_info->domain_attr->mode = 0;
psmx_info->domain_attr->mr_cnt = 65535;
psmx_info->next = NULL;
psmx_info->caps = (hints && hints->caps) ? hints->caps : caps;
psmx_info->mode = mode;
psmx_info->addr_format = FI_ADDR_PSMX;
psmx_info->src_addr = src_addr;
psmx_info->src_addrlen = sizeof(*src_addr);
psmx_info->dest_addr = dest_addr;
psmx_info->dest_addrlen = sizeof(*dest_addr);
psmx_info->fabric_attr->name = strdup(PSMX_FABRIC_NAME);
psmx_info->fabric_attr->prov_name = NULL;
psmx_info->fabric_attr->prov_version = PSMX_VERSION;
psmx_info->tx_attr->caps = psmx_info->caps;
psmx_info->tx_attr->mode = psmx_info->mode;
psmx_info->tx_attr->op_flags = (hints && hints->tx_attr && hints->tx_attr->op_flags)
? hints->tx_attr->op_flags : 0;
psmx_info->tx_attr->msg_order = PSMX_MSG_ORDER;
psmx_info->tx_attr->comp_order = PSMX_COMP_ORDER;
psmx_info->tx_attr->inject_size = PSMX_INJECT_SIZE;
psmx_info->tx_attr->size = UINT64_MAX;
psmx_info->tx_attr->iov_limit = 1;
psmx_info->tx_attr->rma_iov_limit = 1;
psmx_info->rx_attr->caps = psmx_info->caps;
psmx_info->rx_attr->mode = psmx_info->mode;
psmx_info->rx_attr->op_flags = (hints && hints->rx_attr && hints->rx_attr->op_flags)
? hints->rx_attr->op_flags : 0;
psmx_info->rx_attr->msg_order = PSMX_MSG_ORDER;
psmx_info->rx_attr->comp_order = PSMX_COMP_ORDER;
psmx_info->rx_attr->total_buffered_recv = ~(0ULL); /* that's how PSM handles it internally! */
psmx_info->rx_attr->size = UINT64_MAX;
psmx_info->rx_attr->iov_limit = 1;
*info = psmx_info;
return 0;
err_out:
free(dest_addr);
free(src_addr);
return err;
}
static void psmx_fini(void)
{
FI_INFO(&psmx_prov, FI_LOG_CORE, "\n");
if (! --psmx_init_count && psmx_lib_initialized) {
/* This function is called from a library destructor, which is called
* automatically when exit() is called. The call to psm_finalize()
* might cause deadlock if the applicaiton is terminated with Ctrl-C
* -- the application could be inside a PSM call, holding a lock that
* psm_finalize() tries to acquire. This can be avoided by only
* calling psm_finalize() when PSM is guaranteed to be unused.
*/
if (psmx_active_fabric) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"psmx_active_fabric != NULL, skip psm_finalize\n");
} else {
psm_finalize();
psmx_lib_initialized = 0;
}
}
}
static int psmx_getinfo(uint32_t version, const char *node, const char *service,
uint64_t flags, struct fi_info *hints, struct fi_info **info)
{
struct fi_info *psmx_info;
uint32_t cnt = 0;
psm_epid_t *dest_addr = NULL;
struct psmx_src_name *src_addr;
int ep_type = FI_EP_RDM;
int av_type = FI_AV_UNSPEC;
uint64_t mode = FI_CONTEXT;
enum fi_mr_mode mr_mode = FI_MR_SCALABLE;
enum fi_threading threading = FI_THREAD_COMPLETION;
enum fi_progress control_progress = FI_PROGRESS_MANUAL;
enum fi_progress data_progress = FI_PROGRESS_MANUAL;
int caps = 0;
uint64_t max_tag_value = 0;
int err = -FI_ENODATA;
int svc0, svc = PSMX_ANY_SERVICE;
FI_INFO(&psmx_prov, FI_LOG_CORE,"\n");
*info = NULL;
if (psmx_init_lib())
return -FI_ENODATA;
if (psm_ep_num_devunits(&cnt) || !cnt) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"no PSM device is found.\n");
return -FI_ENODATA;
}
psmx_init_env();
src_addr = calloc(1, sizeof(*src_addr));
if (!src_addr) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"failed to allocate src addr.\n");
return -FI_ENODATA;
}
src_addr->unit = PSMX_DEFAULT_UNIT;
src_addr->port = PSMX_DEFAULT_PORT;
src_addr->service = PSMX_ANY_SERVICE;
if (flags & FI_SOURCE) {
if (node)
sscanf(node, "%*[^:]:%d:%d", &src_addr->unit, &src_addr->port);
if (service)
sscanf(service, "%d", &src_addr->service);
FI_INFO(&psmx_prov, FI_LOG_CORE,
"node '%s' service '%s' converted to <unit=%d, port=%d, service=%d>\n",
node, service, src_addr->unit, src_addr->port, src_addr->service);
} else if (node) {
if (service)
svc = atoi(service);
svc0 = svc;
dest_addr = psmx_ns_resolve_name(node, &svc);
if (dest_addr) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"'%s:%u' resolved to <epid=0x%llx>:%u\n", node, svc0,
*dest_addr, svc);
} else {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"failed to resolve '%s:%u'.\n", node, svc);
err = -FI_ENODATA;
goto err_out;
}
}
if (hints) {
switch (hints->addr_format) {
case FI_FORMAT_UNSPEC:
case FI_ADDR_PSMX:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->addr_format=%d, supported=%d,%d.\n",
hints->addr_format, FI_FORMAT_UNSPEC, FI_ADDR_PSMX);
goto err_out;
}
if (hints->ep_attr) {
switch (hints->ep_attr->type) {
case FI_EP_UNSPEC:
case FI_EP_DGRAM:
case FI_EP_RDM:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->type=%d, supported=%d,%d,%d.\n",
hints->ep_attr->type, FI_EP_UNSPEC,
FI_EP_DGRAM, FI_EP_RDM);
goto err_out;
}
switch (hints->ep_attr->protocol) {
case FI_PROTO_UNSPEC:
case FI_PROTO_PSMX:
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->protocol=%d, supported=%d %d\n",
hints->ep_attr->protocol,
FI_PROTO_UNSPEC, FI_PROTO_PSMX);
goto err_out;
}
if (hints->ep_attr->tx_ctx_cnt > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->tx_ctx_cnt=%d, supported=0,1\n",
hints->ep_attr->tx_ctx_cnt);
goto err_out;
}
if (hints->ep_attr->rx_ctx_cnt > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->rx_ctx_cnt=%d, supported=0,1\n",
hints->ep_attr->rx_ctx_cnt);
goto err_out;
}
}
if ((hints->caps & PSMX_CAPS) != hints->caps &&
(hints->caps & PSMX_CAPS2) != hints->caps) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->caps=0x%llx, supported=0x%llx,0x%llx\n",
hints->caps, PSMX_CAPS, PSMX_CAPS2);
goto err_out;
}
if (hints->tx_attr) {
if ((hints->tx_attr->op_flags & PSMX_OP_FLAGS) !=
hints->tx_attr->op_flags) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx->flags=0x%llx, "
"supported=0x%llx\n",
hints->tx_attr->op_flags,
PSMX_OP_FLAGS);
goto err_out;
}
if (hints->tx_attr->inject_size > PSMX_INJECT_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->inject_size=%ld,"
"supported=%ld.\n",
hints->tx_attr->inject_size,
PSMX_INJECT_SIZE);
goto err_out;
}
}
if (hints->rx_attr &&
(hints->rx_attr->op_flags & PSMX_OP_FLAGS) !=
hints->rx_attr->op_flags) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->rx->flags=0x%llx, supported=0x%llx\n",
hints->rx_attr->op_flags, PSMX_OP_FLAGS);
goto err_out;
}
if ((hints->caps & FI_TAGGED) ||
((hints->caps & FI_MSG) && !psmx_env.am_msg)) {
if ((hints->mode & FI_CONTEXT) != FI_CONTEXT) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->mode=0x%llx, required=0x%llx\n",
hints->mode, FI_CONTEXT);
goto err_out;
}
} else {
mode = 0;
}
if (hints->fabric_attr && hints->fabric_attr->name &&
strcmp(hints->fabric_attr->name, PSMX_FABRIC_NAME)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->fabric_name=%s, supported=psm\n",
hints->fabric_attr->name);
goto err_out;
}
if (hints->domain_attr) {
if (hints->domain_attr->name &&
strcmp(hints->domain_attr->name, PSMX_DOMAIN_NAME)) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_name=%s, supported=psm\n",
hints->domain_attr->name);
goto err_out;
}
switch (hints->domain_attr->av_type) {
case FI_AV_UNSPEC:
case FI_AV_MAP:
case FI_AV_TABLE:
av_type = hints->domain_attr->av_type;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->av_type=%d, supported=%d %d %d\n",
hints->domain_attr->av_type, FI_AV_UNSPEC, FI_AV_MAP,
FI_AV_TABLE);
goto err_out;
}
switch (hints->domain_attr->mr_mode) {
case FI_MR_UNSPEC:
break;
case FI_MR_BASIC:
case FI_MR_SCALABLE:
mr_mode = hints->domain_attr->mr_mode;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->mr_mode=%d, supported=%d %d %d\n",
hints->domain_attr->mr_mode, FI_MR_UNSPEC, FI_MR_BASIC,
FI_MR_SCALABLE);
goto err_out;
}
switch (hints->domain_attr->threading) {
case FI_THREAD_UNSPEC:
break;
case FI_THREAD_FID:
case FI_THREAD_ENDPOINT:
case FI_THREAD_COMPLETION:
case FI_THREAD_DOMAIN:
threading = hints->domain_attr->threading;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->threading=%d, supported=%d %d %d %d %d\n",
hints->domain_attr->threading, FI_THREAD_UNSPEC,
FI_THREAD_FID, FI_THREAD_ENDPOINT, FI_THREAD_COMPLETION,
FI_THREAD_DOMAIN);
goto err_out;
}
switch (hints->domain_attr->control_progress) {
case FI_PROGRESS_UNSPEC:
break;
case FI_PROGRESS_MANUAL:
case FI_PROGRESS_AUTO:
control_progress = hints->domain_attr->control_progress;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->control_progress=%d, supported=%d %d %d\n",
hints->domain_attr->control_progress, FI_PROGRESS_UNSPEC,
FI_PROGRESS_MANUAL, FI_PROGRESS_AUTO);
goto err_out;
}
switch (hints->domain_attr->data_progress) {
case FI_PROGRESS_UNSPEC:
break;
case FI_PROGRESS_MANUAL:
case FI_PROGRESS_AUTO:
data_progress = hints->domain_attr->data_progress;
break;
default:
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->data_progress=%d, supported=%d %d %d\n",
hints->domain_attr->data_progress, FI_PROGRESS_UNSPEC,
FI_PROGRESS_MANUAL, FI_PROGRESS_AUTO);
goto err_out;
}
if (hints->domain_attr->caps & FI_SHARED_AV) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->domain_attr->caps=%lx, shared AV is unsupported\n",
hints->domain_attr->caps);
goto err_out;
}
}
if (hints->ep_attr) {
if (hints->ep_attr->max_msg_size > PSMX_MAX_MSG_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->ep_attr->max_msg_size=%ld,"
"supported=%ld.\n",
hints->ep_attr->max_msg_size,
PSMX_MAX_MSG_SIZE);
goto err_out;
}
max_tag_value = fi_tag_bits(hints->ep_attr->mem_tag_format);
}
if (hints->tx_attr) {
if ((hints->tx_attr->msg_order & PSMX_MSG_ORDER) !=
hints->tx_attr->msg_order) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->msg_order=%lx,"
"supported=%lx.\n",
hints->tx_attr->msg_order,
PSMX_MSG_ORDER);
goto err_out;
}
if ((hints->tx_attr->comp_order & PSMX_COMP_ORDER) !=
hints->tx_attr->comp_order) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->msg_order=%lx,"
"supported=%lx.\n",
hints->tx_attr->comp_order,
PSMX_COMP_ORDER);
goto err_out;
}
if (hints->tx_attr->inject_size > PSMX_INJECT_SIZE) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->inject_size=%ld,"
"supported=%d.\n",
hints->tx_attr->inject_size,
PSMX_INJECT_SIZE);
goto err_out;
}
if (hints->tx_attr->iov_limit > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->iov_limit=%ld,"
"supported=1.\n",
hints->tx_attr->iov_limit);
goto err_out;
}
if (hints->tx_attr->rma_iov_limit > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->tx_attr->rma_iov_limit=%ld,"
"supported=1.\n",
hints->tx_attr->rma_iov_limit);
goto err_out;
}
}
if (hints->rx_attr) {
if ((hints->rx_attr->msg_order & PSMX_MSG_ORDER) !=
hints->rx_attr->msg_order) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->rx_attr->msg_order=%lx,"
"supported=%lx.\n",
hints->rx_attr->msg_order,
PSMX_MSG_ORDER);
goto err_out;
}
if ((hints->rx_attr->comp_order & PSMX_COMP_ORDER) !=
hints->rx_attr->comp_order) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->rx_attr->msg_order=%lx,"
"supported=%lx.\n",
hints->rx_attr->comp_order,
PSMX_COMP_ORDER);
goto err_out;
}
if (hints->rx_attr->iov_limit > 1) {
FI_INFO(&psmx_prov, FI_LOG_CORE,
"hints->rx_attr->iov_limit=%ld,"
"supported=1.\n",
hints->rx_attr->iov_limit);
goto err_out;
}
}
caps = hints->caps;
/* TODO: check other fields of hints */
}
if (psmx_reserve_tag_bits(&caps, &max_tag_value) < 0)
goto err_out;
psmx_info = fi_allocinfo();
if (!psmx_info) {
err = -FI_ENOMEM;
goto err_out;
}
psmx_info->ep_attr->type = ep_type;
psmx_info->ep_attr->protocol = FI_PROTO_PSMX;
psmx_info->ep_attr->protocol_version = PSM_VERNO;
psmx_info->ep_attr->max_msg_size = PSMX_MAX_MSG_SIZE;
psmx_info->ep_attr->mem_tag_format = fi_tag_format(max_tag_value);
psmx_info->ep_attr->tx_ctx_cnt = 1;
psmx_info->ep_attr->rx_ctx_cnt = 1;
psmx_info->domain_attr->threading = threading;
psmx_info->domain_attr->control_progress = control_progress;
psmx_info->domain_attr->data_progress = data_progress;
psmx_info->domain_attr->name = strdup(PSMX_DOMAIN_NAME);
psmx_info->domain_attr->resource_mgmt = FI_RM_ENABLED;
psmx_info->domain_attr->av_type = av_type;
psmx_info->domain_attr->mr_mode = mr_mode;
psmx_info->domain_attr->mr_key_size = sizeof(uint64_t);
psmx_info->domain_attr->cq_data_size = 4;
psmx_info->domain_attr->cq_cnt = 65535;
psmx_info->domain_attr->ep_cnt = 65535;
psmx_info->domain_attr->tx_ctx_cnt = 1;
psmx_info->domain_attr->rx_ctx_cnt = 1;
psmx_info->domain_attr->max_ep_tx_ctx = 1;
psmx_info->domain_attr->max_ep_rx_ctx = 1;
psmx_info->domain_attr->max_ep_stx_ctx = 65535;
psmx_info->domain_attr->max_ep_srx_ctx = 0;
psmx_info->domain_attr->cntr_cnt = 65535;
psmx_info->domain_attr->mr_iov_limit = 65535;
psmx_info->domain_attr->caps = PSMX_DOM_CAPS;
psmx_info->domain_attr->mode = 0;
psmx_info->next = NULL;
psmx_info->caps = (hints && hints->caps) ? hints->caps : caps;
psmx_info->mode = mode;
psmx_info->addr_format = FI_ADDR_PSMX;
psmx_info->src_addr = src_addr;
psmx_info->src_addrlen = sizeof(*src_addr);
psmx_info->dest_addr = dest_addr;
psmx_info->dest_addrlen = sizeof(*dest_addr);
psmx_info->fabric_attr->name = strdup(PSMX_FABRIC_NAME);
psmx_info->fabric_attr->prov_name = NULL;
psmx_info->fabric_attr->prov_version = PSMX_VERSION;
psmx_info->tx_attr->caps = psmx_info->caps;
psmx_info->tx_attr->mode = psmx_info->mode;
psmx_info->tx_attr->op_flags = (hints && hints->tx_attr && hints->tx_attr->op_flags)
? hints->tx_attr->op_flags : 0;
psmx_info->tx_attr->msg_order = PSMX_MSG_ORDER;
psmx_info->tx_attr->comp_order = PSMX_COMP_ORDER;
psmx_info->tx_attr->inject_size = PSMX_INJECT_SIZE;
psmx_info->tx_attr->size = UINT64_MAX;
psmx_info->tx_attr->iov_limit = 1;
psmx_info->tx_attr->rma_iov_limit = 1;
psmx_info->rx_attr->caps = psmx_info->caps;
psmx_info->rx_attr->mode = psmx_info->mode;
psmx_info->rx_attr->op_flags = (hints && hints->rx_attr && hints->rx_attr->op_flags)
? hints->rx_attr->op_flags : 0;
psmx_info->rx_attr->msg_order = PSMX_MSG_ORDER;
psmx_info->rx_attr->comp_order = PSMX_COMP_ORDER;
psmx_info->rx_attr->total_buffered_recv = ~(0ULL); /* that's how PSM handles it internally! */
psmx_info->rx_attr->size = UINT64_MAX;
psmx_info->rx_attr->iov_limit = 1;
*info = psmx_info;
return 0;
err_out:
free(dest_addr);
free(src_addr);
return err;
}
